More GIProbe work and fixes

10 files changed, 98 insertions, 36 deletions

diff --git a/servers/visual/rasterizer_rd/rasterizer_rd.cpp b/servers/visual/rasterizer_rd/rasterizer_rd.cpp
index fc38b90b88..89465b580c 100644
--- a/servers/visual/rasterizer_rd/rasterizer_rd.cpp
+++ b/servers/visual/rasterizer_rd/rasterizer_rd.cpp
@@ -84,12 +84,10 @@ void RasterizerRD::begin_frame(double frame_step) {
 
 void RasterizerRD::end_frame(bool p_swap_buffers) {
 
-	RD::get_singleton()->finalize_frame();
 #ifndef _MSC_VER
-#warning not swapping buffers likely not an option for now, find another way
+#warning TODO: likely passa bool to swap buffers to avoid display?
 #endif
-	OS::get_singleton()->swap_buffers(); //probably should pass some bool to avoid display?
-	RD::get_singleton()->advance_frame(); //advance frame here, so any new call happens on new frame
+	RD::get_singleton()->swap_buffers(); //probably should pass some bool to avoid display?
 }
 
 void RasterizerRD::initialize() {
diff --git a/servers/visual/rasterizer_rd/rasterizer_scene_forward_rd.cpp b/servers/visual/rasterizer_rd/rasterizer_scene_forward_rd.cpp
index 502ac9196f..55974846f2 100644
--- a/servers/visual/rasterizer_rd/rasterizer_scene_forward_rd.cpp
+++ b/servers/visual/rasterizer_rd/rasterizer_scene_forward_rd.cpp
@@ -1899,7 +1899,7 @@ void RasterizerSceneForwardRD::_render_scene(RenderBufferData *p_buffer_data, co
 	_fill_instances(render_list.elements, render_list.element_count);
 
 	bool can_continue = true; //unless the middle buffers are needed
-	bool debug_giprobes = debug_draw == VS::VIEWPORT_DEBUG_DRAW_GI_PROBE_ALBEDO || debug_draw == VS::VIEWPORT_DEBUG_DRAW_GI_PROBE_LIGHTING;
+	bool debug_giprobes = debug_draw == VS::VIEWPORT_DEBUG_DRAW_GI_PROBE_ALBEDO || debug_draw == VS::VIEWPORT_DEBUG_DRAW_GI_PROBE_LIGHTING || debug_draw == VS::VIEWPORT_DEBUG_DRAW_GI_PROBE_EMISSION;
 	bool using_separate_specular = false;
 
 	bool depth_pre_pass = depth_framebuffer.is_valid();
@@ -1930,7 +1930,7 @@ void RasterizerSceneForwardRD::_render_scene(RenderBufferData *p_buffer_data, co
 		CameraMatrix cm = (dc * p_cam_projection) * CameraMatrix(p_cam_transform.affine_inverse());
 		RD::DrawListID draw_list = RD::get_singleton()->draw_list_begin(opaque_framebuffer, RD::INITIAL_ACTION_CONTINUE, will_continue ? RD::FINAL_ACTION_CONTINUE : RD::FINAL_ACTION_READ, RD::INITIAL_ACTION_CONTINUE, will_continue ? RD::FINAL_ACTION_CONTINUE : RD::FINAL_ACTION_READ);
 		for (int i = 0; i < p_gi_probe_cull_count; i++) {
-			_debug_giprobe(p_gi_probe_cull_result[i], draw_list, opaque_framebuffer, cm, debug_draw == VS::VIEWPORT_DEBUG_DRAW_GI_PROBE_LIGHTING, 1.0);
+			_debug_giprobe(p_gi_probe_cull_result[i], draw_list, opaque_framebuffer, cm, debug_draw == VS::VIEWPORT_DEBUG_DRAW_GI_PROBE_LIGHTING, debug_draw == VS::VIEWPORT_DEBUG_DRAW_GI_PROBE_EMISSION, 1.0);
 		}
 		RD::get_singleton()->draw_list_end();
 	}
@@ -2204,10 +2204,15 @@ void RasterizerSceneForwardRD::_update_render_base_uniform_set() {
 			} else {
 				u.ids.resize(slot_count);
 			}
+
+			print_line("updating slots, probe count: " + itos(slot_count));
 			for (int i = 0; i < slot_count; i++) {
 
 				RID probe = gi_probe_get_slots()[i];
 
+				if (probe.is_valid()) {
+					print_line("probe valid: " + itos(i));
+				}
 				if (gi_probe_is_anisotropic()) {
 					if (probe.is_null()) {
 						RID empty_tex = storage->texture_rd_get_default(RasterizerStorageRD::DEFAULT_RD_TEXTURE_3D_WHITE);
diff --git a/servers/visual/rasterizer_rd/rasterizer_scene_rd.cpp b/servers/visual/rasterizer_rd/rasterizer_scene_rd.cpp
index 1b7d5a8e47..f039bae766 100644
--- a/servers/visual/rasterizer_rd/rasterizer_scene_rd.cpp
+++ b/servers/visual/rasterizer_rd/rasterizer_scene_rd.cpp
@@ -1439,6 +1439,7 @@ void RasterizerSceneRD::gi_probe_update(RID p_probe, const Vector<RID> &p_light_
 		}
 
 		gi_probe->last_probe_data_version = data_version;
+		gi_probe_slots_dirty = true;
 	}
 
 	// UDPDATE TIME
@@ -1510,6 +1511,8 @@ void RasterizerSceneRD::gi_probe_update(RID p_probe, const Vector<RID> &p_light_
 		RD::ComputeListID compute_list = RD::get_singleton()->compute_list_begin();
 
 		int passes = storage->gi_probe_is_using_two_bounces(gi_probe->probe) ? 2 : 1;
+		int wg_size = 64;
+		int wg_limit_x = RD::get_singleton()->limit_get(RD::LIMIT_MAX_COMPUTE_WORKGROUP_COUNT_X);
 
 		for (int pass = 0; pass < passes; pass++) {
 
@@ -1532,9 +1535,14 @@ void RasterizerSceneRD::gi_probe_update(RID p_probe, const Vector<RID> &p_light_
 				push_constant.cell_offset = gi_probe->mipmaps[i].cell_offset;
 				push_constant.cell_count = gi_probe->mipmaps[i].cell_count;
 
-				RD::get_singleton()->compute_list_set_push_constant(compute_list, &push_constant, sizeof(GIProbePushConstant));
-
-				RD::get_singleton()->compute_list_dispatch(compute_list, (gi_probe->mipmaps[i].cell_count - 1) / 64 + 1, 1, 1);
+				int wg_todo = (gi_probe->mipmaps[i].cell_count - 1) / wg_size + 1;
+				while (wg_todo) {
+					int wg_count = MIN(wg_todo, wg_limit_x);
+					RD::get_singleton()->compute_list_set_push_constant(compute_list, &push_constant, sizeof(GIProbePushConstant));
+					RD::get_singleton()->compute_list_dispatch(compute_list, wg_count, 1, 1);
+					wg_todo -= wg_count;
+					push_constant.cell_offset += wg_count * wg_size;
+				}
 			}
 
 			RD::get_singleton()->compute_list_add_barrier(compute_list); //wait til previous step is done
@@ -1548,9 +1556,14 @@ void RasterizerSceneRD::gi_probe_update(RID p_probe, const Vector<RID> &p_light_
 				push_constant.cell_offset = gi_probe->mipmaps[i].cell_offset;
 				push_constant.cell_count = gi_probe->mipmaps[i].cell_count;
 
-				RD::get_singleton()->compute_list_set_push_constant(compute_list, &push_constant, sizeof(GIProbePushConstant));
-
-				RD::get_singleton()->compute_list_dispatch(compute_list, (gi_probe->mipmaps[i].cell_count - 1) / 64 + 1, 1, 1);
+				int wg_todo = (gi_probe->mipmaps[i].cell_count - 1) / wg_size + 1;
+				while (wg_todo) {
+					int wg_count = MIN(wg_todo, wg_limit_x);
+					RD::get_singleton()->compute_list_set_push_constant(compute_list, &push_constant, sizeof(GIProbePushConstant));
+					RD::get_singleton()->compute_list_dispatch(compute_list, wg_count, 1, 1);
+					wg_todo -= wg_count;
+					push_constant.cell_offset += wg_count * wg_size;
+				}
 			}
 		}
 
@@ -1560,7 +1573,7 @@ void RasterizerSceneRD::gi_probe_update(RID p_probe, const Vector<RID> &p_light_
 	gi_probe->last_probe_version = storage->gi_probe_get_version(gi_probe->probe);
 }
 
-void RasterizerSceneRD::_debug_giprobe(RID p_gi_probe, RD::DrawListID p_draw_list, RID p_framebuffer, const CameraMatrix &p_camera_with_transform, bool p_lighting, float p_alpha) {
+void RasterizerSceneRD::_debug_giprobe(RID p_gi_probe, RD::DrawListID p_draw_list, RID p_framebuffer, const CameraMatrix &p_camera_with_transform, bool p_lighting, bool p_emission, float p_alpha) {
 	GIProbeInstance *gi_probe = gi_probe_instance_owner.getornull(p_gi_probe);
 	ERR_FAIL_COND(!gi_probe);
 
@@ -1631,7 +1644,7 @@ void RasterizerSceneRD::_debug_giprobe(RID p_gi_probe, RD::DrawListID p_draw_lis
 	}
 
 	giprobe_debug_uniform_set = RD::get_singleton()->uniform_set_create(uniforms, giprobe_debug_shader_version_shaders[0], 0);
-	RD::get_singleton()->draw_list_bind_render_pipeline(p_draw_list, giprobe_debug_shader_version_pipelines[p_lighting ? GI_PROBE_DEBUG_LIGHT : GI_PROBE_DEBUG_COLOR].get_render_pipeline(RD::INVALID_ID, RD::get_singleton()->framebuffer_get_format(p_framebuffer)));
+	RD::get_singleton()->draw_list_bind_render_pipeline(p_draw_list, giprobe_debug_shader_version_pipelines[p_emission ? GI_PROBE_DEBUG_EMISSION : p_lighting ? GI_PROBE_DEBUG_LIGHT : GI_PROBE_DEBUG_COLOR].get_render_pipeline(RD::INVALID_ID, RD::get_singleton()->framebuffer_get_format(p_framebuffer)));
 	RD::get_singleton()->draw_list_bind_uniform_set(p_draw_list, giprobe_debug_uniform_set, 0);
 	RD::get_singleton()->draw_list_set_push_constant(p_draw_list, &push_constant, sizeof(GIProbeDebugPushConstant));
 	RD::get_singleton()->draw_list_draw(p_draw_list, false, cell_count, 36);
@@ -2007,6 +2020,7 @@ RasterizerSceneRD::RasterizerSceneRD(RasterizerStorageRD *p_storage) {
 		Vector<String> versions;
 		versions.push_back("\n#define MODE_DEBUG_COLOR\n");
 		versions.push_back("\n#define MODE_DEBUG_LIGHT\n");
+		versions.push_back("\n#define MODE_DEBUG_EMISSION\n");
 
 		giprobe_debug_shader.initialize(versions, defines);
 		giprobe_debug_shader_version = giprobe_debug_shader.version_create();
diff --git a/servers/visual/rasterizer_rd/rasterizer_scene_rd.h b/servers/visual/rasterizer_rd/rasterizer_scene_rd.h
index 7acb66c587..d05c9a9328 100644
--- a/servers/visual/rasterizer_rd/rasterizer_scene_rd.h
+++ b/servers/visual/rasterizer_rd/rasterizer_scene_rd.h
@@ -27,7 +27,7 @@ protected:
 	virtual void _render_scene(RenderBufferData *p_buffer_data, const Transform &p_cam_transform, const CameraMatrix &p_cam_projection, bool p_cam_ortogonal, InstanceBase **p_cull_result, int p_cull_count, RID *p_light_cull_result, int p_light_cull_count, RID *p_reflection_probe_cull_result, int p_reflection_probe_cull_count, RID *p_gi_probe_cull_result, int p_gi_probe_cull_count, RID p_environment, RID p_shadow_atlas, RID p_reflection_atlas, RID p_reflection_probe, int p_reflection_probe_pass) = 0;
 	virtual void _render_shadow(RID p_framebuffer, InstanceBase **p_cull_result, int p_cull_count, const CameraMatrix &p_projection, const Transform &p_transform, float p_zfar, float p_bias, float p_normal_bias, bool p_use_dp, bool use_dp_flip) = 0;
 
-	virtual void _debug_giprobe(RID p_gi_probe, RenderingDevice::DrawListID p_draw_list, RID p_framebuffer, const CameraMatrix &p_camera_with_transform, bool p_lighting, float p_alpha);
+	virtual void _debug_giprobe(RID p_gi_probe, RenderingDevice::DrawListID p_draw_list, RID p_framebuffer, const CameraMatrix &p_camera_with_transform, bool p_lighting, bool p_emission, float p_alpha);
 
 private:
 	int roughness_layers;
@@ -210,6 +210,7 @@ private:
 	enum {
 		GI_PROBE_DEBUG_COLOR,
 		GI_PROBE_DEBUG_LIGHT,
+		GI_PROBE_DEBUG_EMISSION,
 		GI_PROBE_DEBUG_MAX
 	};
 
diff --git a/servers/visual/rasterizer_rd/shaders/giprobe.glsl b/servers/visual/rasterizer_rd/shaders/giprobe.glsl
index 35b8d6ba6b..a723490e8b 100644
--- a/servers/visual/rasterizer_rd/shaders/giprobe.glsl
+++ b/servers/visual/rasterizer_rd/shaders/giprobe.glsl
@@ -248,7 +248,7 @@ void main() {
 
 	vec3 pos = vec3(posu) + vec3(0.5);
 
-	vec3 emission = vec3(ivec3(cell_data.data[cell_index].emission&0x3FF,(cell_data.data[cell_index].emission>>10)&0x7FF,cell_data.data[cell_index].emission>>21)) * params.emission_scale;
+	vec3 emission = vec3(uvec3(cell_data.data[cell_index].emission & 0x1ff,(cell_data.data[cell_index].emission >> 9) & 0x1ff,(cell_data.data[cell_index].emission >> 18) & 0x1ff)) * pow(2.0, float(cell_data.data[cell_index].emission >> 27) - 15.0 - 9.0);
 	vec4 normal = unpackSnorm4x8(cell_data.data[cell_index].normal);
 
 #ifdef MODE_ANISOTROPIC
@@ -271,8 +271,8 @@ void main() {
 		float distance = length(light_dir);
 		light_dir=normalize(light_dir);
 
-		if (length(normal.xyz) > 0.2 && dot(normal.xyz,light_dir)>=0) {
-			continue; //not facing the light
+		if (attenuation < 0.01 || (length(normal.xyz) > 0.2 && dot(normal.xyz,light_dir)>=0)) {
+			continue; //not facing the light, or attenuation is near zero
 		}
 
 		if (lights.data[i].has_shadow) {
@@ -299,11 +299,11 @@ void main() {
 #ifdef MODE_ANISOTROPIC
 		for(uint j=0;j<6;j++) {
 
-			accum[j]+=max(0.0,dot(accum_dirs[j],-light_dir))*light+emission;
+			accum[j]+=max(0.0,dot(accum_dirs[j],-light_dir))*light;
 		}
 #else
 		if (length(normal.xyz) > 0.2) {
-			accum+=max(0.0,dot(normal.xyz,-light_dir))*light+emission;
+			accum+=max(0.0,dot(normal.xyz,-light_dir))*light;
 		} else {
 			//all directions
 			accum+=light+emission;
@@ -314,14 +314,14 @@ void main() {
 
 #ifdef MODE_ANISOTROPIC
 
-	outputs.data[cell_index*6+0]=vec4(accum[0],0.0);
-	outputs.data[cell_index*6+1]=vec4(accum[1],0.0);
-	outputs.data[cell_index*6+2]=vec4(accum[2],0.0);
-	outputs.data[cell_index*6+3]=vec4(accum[3],0.0);
-	outputs.data[cell_index*6+4]=vec4(accum[4],0.0);
-	outputs.data[cell_index*6+5]=vec4(accum[5],0.0);
+	outputs.data[cell_index*6+0]=vec4(accum[0] + emission,0.0);
+	outputs.data[cell_index*6+1]=vec4(accum[1] + emission,0.0);
+	outputs.data[cell_index*6+2]=vec4(accum[2] + emission,0.0);
+	outputs.data[cell_index*6+3]=vec4(accum[3] + emission,0.0);
+	outputs.data[cell_index*6+4]=vec4(accum[4] + emission,0.0);
+	outputs.data[cell_index*6+5]=vec4(accum[5] + emission,0.0);
 #else
-	outputs.data[cell_index]=vec4(accum,0.0);
+	outputs.data[cell_index]=vec4(accum + emission,0.0);
 
 #endif
 
@@ -420,7 +420,7 @@ void main() {
 				}
 
 			}
-			color *= cone_weights[i] * params.dynamic_range; //restore range
+			color *= cone_weights[i] * vec4(albedo.rgb,1.0) * params.dynamic_range; //restore range
 #ifdef MODE_ANISOTROPIC
 			for(uint j=0;j<6;j++) {
 
diff --git a/servers/visual/rasterizer_rd/shaders/giprobe_debug.glsl b/servers/visual/rasterizer_rd/shaders/giprobe_debug.glsl
index deaeb771b9..71ecaffde7 100644
--- a/servers/visual/rasterizer_rd/shaders/giprobe_debug.glsl
+++ b/servers/visual/rasterizer_rd/shaders/giprobe_debug.glsl
@@ -85,9 +85,14 @@ void main() {
 
 	uvec3 posu = uvec3(cell_data.data[cell_index].position&0x7FF,(cell_data.data[cell_index].position>>11)&0x3FF,cell_data.data[cell_index].position>>21);
 
+#ifdef MODE_DEBUG_EMISSION
+	color_interp.xyz = vec3(uvec3(cell_data.data[cell_index].emission & 0x1ff,(cell_data.data[cell_index].emission >> 9) & 0x1ff,(cell_data.data[cell_index].emission >> 18) & 0x1ff)) * pow(2.0, float(cell_data.data[cell_index].emission >> 27) - 15.0 - 9.0);
+#endif
+
 #ifdef MODE_DEBUG_COLOR
 	color_interp.xyz = unpackUnorm4x8(cell_data.data[cell_index].albedo).xyz;
 #endif
+
 #ifdef MODE_DEBUG_LIGHT
 
 #ifdef USE_ANISOTROPY
diff --git a/servers/visual/rasterizer_rd/shaders/scene_forward.glsl b/servers/visual/rasterizer_rd/shaders/scene_forward.glsl
index 2070df3952..a54a84536a 100644
--- a/servers/visual/rasterizer_rd/shaders/scene_forward.glsl
+++ b/servers/visual/rasterizer_rd/shaders/scene_forward.glsl
@@ -326,9 +326,9 @@ layout(location =8) in float dp_clip;
 
 //defines to keep compatibility with vertex
 
-#define world_matrix instances.data[instance_index].transform;
-#define world_normal_matrix instances.data[instance_index].normal_transform;
-#define projection_matrix scene_data.projection_matrix;
+#define world_matrix instances.data[instance_index].transform
+#define world_normal_matrix instances.data[instance_index].normal_transform
+#define projection_matrix scene_data.projection_matrix
 
 #ifdef USE_MATERIAL_UNIFORMS
 layout(set = 3, binding = 0, std140) uniform MaterialUniforms {
diff --git a/servers/visual/rendering_device.h b/servers/visual/rendering_device.h
index f5a212b076..75eaab796a 100644
--- a/servers/visual/rendering_device.h
+++ b/servers/visual/rendering_device.h
@@ -1002,14 +1002,23 @@ public:
 		LIMIT_MAX_VERTEX_INPUT_BINDINGS,
 		LIMIT_MAX_VERTEX_INPUT_BINDING_STRIDE,
 		LIMIT_MIN_UNIFORM_BUFFER_OFFSET_ALIGNMENT,
+		LIMIT_MAX_COMPUTE_SHARED_MEMORY_SIZE,
+		LIMIT_MAX_COMPUTE_WORKGROUP_COUNT_X,
+		LIMIT_MAX_COMPUTE_WORKGROUP_COUNT_Y,
+		LIMIT_MAX_COMPUTE_WORKGROUP_COUNT_Z,
+		LIMIT_MAX_COMPUTE_WORKGROUP_INVOCATIONS,
+		LIMIT_MAX_COMPUTE_WORKGROUP_SIZE_X,
+		LIMIT_MAX_COMPUTE_WORKGROUP_SIZE_Y,
+		LIMIT_MAX_COMPUTE_WORKGROUP_SIZE_Z,
 	};
 
 	virtual int limit_get(Limit p_limit) = 0;
 
 	//methods below not exposed, used by RenderingDeviceRD
 	virtual void prepare_screen_for_drawing() = 0;
-	virtual void finalize_frame() = 0;
-	virtual void advance_frame() = 0;
+
+	virtual void swap_buffers() = 0;
+
 	virtual uint32_t get_frame_delay() const = 0;
 
 	static RenderingDevice *get_singleton();
diff --git a/servers/visual/visual_server_scene.cpp b/servers/visual/visual_server_scene.cpp
index f7b118bfe5..8d38cf02b3 100644
--- a/servers/visual/visual_server_scene.cpp
+++ b/servers/visual/visual_server_scene.cpp
@@ -345,7 +345,11 @@ void VisualServerScene::instance_set_base(RID p_instance, RID p_base) {
 			case VS::INSTANCE_LIGHT: {
 
 				InstanceLightData *light = static_cast<InstanceLightData *>(instance->base_data);
-
+#ifdef DEBUG_ENABLED
+				if (light->geometries.size()) {
+					ERR_PRINT("BUG, indexing did not unpair geometries from light.");
+				}
+#endif
 				if (instance->scenario && light->D) {
 					instance->scenario->directional_lights.erase(light->D);
 					light->D = NULL;
@@ -371,7 +375,16 @@ void VisualServerScene::instance_set_base(RID p_instance, RID p_base) {
 			case VS::INSTANCE_GI_PROBE: {
 
 				InstanceGIProbeData *gi_probe = static_cast<InstanceGIProbeData *>(instance->base_data);
-
+#ifdef DEBUG_ENABLED
+				if (gi_probe->geometries.size()) {
+					ERR_PRINT("BUG, indexing did not unpair geometries from GIProbe.");
+				}
+#endif
+#ifdef DEBUG_ENABLED
+				if (gi_probe->lights.size()) {
+					ERR_PRINT("BUG, indexing did not unpair lights from GIProbe.");
+				}
+#endif
 				if (gi_probe->update_element.in_list()) {
 					gi_probe_update_list.remove(&gi_probe->update_element);
 				}
@@ -490,7 +503,11 @@ void VisualServerScene::instance_set_scenario(RID p_instance, RID p_scenario) {
 			case VS::INSTANCE_LIGHT: {
 
 				InstanceLightData *light = static_cast<InstanceLightData *>(instance->base_data);
-
+#ifdef DEBUG_ENABLED
+				if (light->geometries.size()) {
+					ERR_PRINT("BUG, indexing did not unpair geometries from light.");
+				}
+#endif
 				if (light->D) {
 					instance->scenario->directional_lights.erase(light->D);
 					light->D = NULL;
@@ -504,6 +521,18 @@ void VisualServerScene::instance_set_scenario(RID p_instance, RID p_scenario) {
 			case VS::INSTANCE_GI_PROBE: {
 
 				InstanceGIProbeData *gi_probe = static_cast<InstanceGIProbeData *>(instance->base_data);
+
+#ifdef DEBUG_ENABLED
+				if (gi_probe->geometries.size()) {
+					ERR_PRINT("BUG, indexing did not unpair geometries from GIProbe.");
+				}
+#endif
+#ifdef DEBUG_ENABLED
+				if (gi_probe->lights.size()) {
+					ERR_PRINT("BUG, indexing did not unpair lights from GIProbe.");
+				}
+#endif
+
 				if (gi_probe->update_element.in_list()) {
 					gi_probe_update_list.remove(&gi_probe->update_element);
 				}
diff --git a/servers/visual_server.h b/servers/visual_server.h
index 5408ad665b..3523afc305 100644
--- a/servers/visual_server.h
+++ b/servers/visual_server.h
@@ -652,6 +652,7 @@ public:
 		VIEWPORT_DEBUG_DRAW_WIREFRAME,
 		VIEWPORT_DEBUG_DRAW_GI_PROBE_ALBEDO,
 		VIEWPORT_DEBUG_DRAW_GI_PROBE_LIGHTING,
+		VIEWPORT_DEBUG_DRAW_GI_PROBE_EMISSION,
 		VIEWPORT_DEBUG_DRAW_SHADOW_ATLAS,
 		VIEWPORT_DEBUG_DRAW_DIRECTIONAL_SHADOW_ATLAS,